Unanswered

Copilot and tabled data

Like (0) Share

Report

Posted on 27 Feb 2024 14:47:06 by theMac

I'd like to understand how Copilot reads and prompts to user informations based on data structured in table when Generative AI answers are prompted.
I gave to CP a twelve rows xlsx file as data source (I tried even with a public html page with data tabled but I'd like to avoid this way) with columns like the following:
ProductCode | Description | DeliveryTime| | Stock Existence | RotationIndex |
Excel file contains only text/numeric data without dynamic/calculated values.

I see here https://learn.microsoft.com/it-it/microsoft-copilot-studio/nlu-documents#supported-document-types that Excel and other tabled data formats are allowed but my experience is a bit unhappy: I see that CP is not always able to prompt me the right informations crossing rows and columns so if I ask "can you tell me the stock existence of product XXXXX?" it answer me prompting an information from the wrong column or it tells me he doesn't understand my question...
My question is: have I to format data in a specific way (i.e. avoiding space character in col's name), are there naming rules for columns I'm not aware of, or is there a way more effective to submit data structured to CP to make it more efficient?

Categories:

General

All responses (2)

Answers (0)

theMac on 27 Feb 2024 at 17:14:13

Like (0)

Report

Re: Copilot and tabled data

I had a test trying to use a txt file with rows and pipeline-separated columns...
first attempt worked fine, CP found the correct answer fast and effective, but then failed all other answer
(I even should find a way to hide references when answer is from that file)
Jeanxyz 72 on 27 Feb 2024 at 16:13:14

Like (0)

Report

Re: Copilot and tabled data

I think the problem is by default, openai retrieve text data based similarity search. Words appear next to each other are more likely to be retrieved as an answer. If you load a table, openai data model converts it into a text file like:
invoice id, invoice date, customer,
1000, 01/10/2024, Microsoft,
1001, 01/01/2024, Google.

When you ask questions about invoice 1000, copilot has a good chance to give you the correct answer because the word Microsoft is closer to 1000 than google. But generally speaking copilot is not reliable to read structured data. Microsoft staff says users need to load structured data into a power platform component for copilot to interpret data correctly.